antinucleon
commented
9 years ago Sorry we are going to deprecate cxxnet project and fully transfer to MXNET in next a few days (or next week). We are still working on some final doc and final debug/feature. If you like you make take a look of https://github.com/dmlc/mxnet/blob/master/example/imagenet/README.md
I installed cxxnet successfully and it can run on Mnist dataset. When I tried cxxnet on ImageNet Dataset, it can still run, but it may halt unexpectedly and randomly, sometimes when loading the mean_224.bin, sometimes when training some batches of images, as shown below. And no error prompted, just stayed with no changes in number of batches and time elapsed, and the same status lasts more than several hours. I can only kill the process with Ctrl+C. The memory of the computer is enough with 5GB used by cxxnet and 2GB left free.
I used kaiming.conf as shown below when running cxxnet.
Configuration for ImageNet
Acknowledgement:
Ref: He, Kaiming, and Jian Sun. "Convolutional Neural Networks at Constrained Time Cost." CVPR2015
J' model in the paper above
data = train iter = imgrec
image_rec = "train.bin" image_mean = "mean_224.bin" rand_crop=1 rand_mirror=1 min_crop_size=192 max_crop_size=224 max_aspect_ratio=0.3 iter = threadbuffer iter = end
eval = test iter = imgrec
image_rec = "test.bin" image_mean = "mean_224.bin"
iter = end
Stage 1
netconfig=start layer[0->1] = conv:conv1 kernel_size = 7 stride = 2 nchannel = 64 layer[1->2] = relu:relu1 layer[2->3] = max_pooling kernel_size = 3
Stage 2
layer[3->4] = conv:conv2 nchannel = 128 kernel_size = 2 stride = 3 layer[4->5] = relu:relu2
layer[5->6] = conv:conv3 nchannel = 128 kernel_size = 2 pad = 1 layer[6->7] = relu:relu3
layer[7->8] = conv:conv4 nchannel = 128 kernel_size = 2 layer[8->9] = relu:relu4
layer[9->10] = conv:conv5 nchannel = 128 kernel_size = 2 pad = 1 layer[10->11] = relu:relu5
layer[11->12] = max_pooling:pool1 kernel_size = 3
Stage 3
layer[12->13] = conv:conv6 nchannel = 256 kernel_size = 2 stride = 2 layer[13->14] = relu:relu6
layer[14->15] = conv:conv7 nchannel = 256 kernel_size = 2 pad = 1 layer[15->16] = relu:relu7
layer[16->17] = conv:conv8 nchannel = 256 kernel_size = 2 layer[17->18] = relu:relu8
layer[18->19] = conv:conv9 nchannel = 256 kernel_size = 2 pad = 1 layer[19->20] = relu:relu9
layer[20->21] = max_pooling:pool2 kernel_size = 3
Stage 4
layer[21->22] = conv:conv10 nchannel = 2304 kernel_size = 2 stride = 3 layer[22->23] = relu:relu10
layer[23->24] = conv:conv11 nchannel = 256 kernel_size = 2 pad = 1 layer[24->25] = relu:relu11
Stage 5
layer[25->26,27,28,29] = split:split1 layer[26->30] = max_pooling:pool3 kernel_size = 1 stride = 1 layer[27->31] = max_pooling:pool4 kernel_size = 2 stride = 2 layer[28->32] = max_pooling:pool5 kernel_size = 3 stride = 3 layer[29->33] = max_pooling:pool6 kernel_size = 6 stride = 6
layer[30->34] = flatten:f1 layer[31->35] = flatten:f2 layer[32->36] = flatten:f3 layer[33->37] = flatten:f4 layer[34,35,36,37->38] = concat:concat1
Stage 6
layer[38->39] = fullc:fc1 nhidden = 4096 layer[39->40] = relu:relu12 layer[40->40] = dropout threshold = 0.5
layer[40->41] = fullc:fc2 nhidden = 4096 layer[41->42] = relu:relu13 layer[42->42] = dropout threshold = 0.5
layer[42->43] = fullc:fc3 nhidden = 1000 layer[43->43] = softmax:softmax1 netconfig=end
evaluation metric
metric = rec@1 metric = rec@5
max_round = 100 num_round = 100
input shape not including batch
input_shape = 3,224,224
batch_size = 256
global parameters in any sectiion outside netconfig, and iter
momentum = 0.9 wmat:lr = 0.01 wmat:wd = 0.0005
bias:wd = 0.000 bias:lr = 0.02
all the learning rate schedule starts with lr
lr:schedule = factor lr:gamma = 0.1 lr:step = 300000
save_model=1 model_dir=models print_step=1
random config
random_type = xavier
new line
dev = gpu:0,1
dev = cpu